better softmax clipping

rl_portfolio_management/util.py

@@ -16,7 +16,8 @@ def MDD(returns):
 
 def softmax(w, t=1.0):
     """softmax implemented in numpy."""
-    w = np.clip(w, -1000, 50)  # avoid inf/nan
+    log_eps = np.log(eps)
+    w = np.clip(w, -log_eps, log_eps)  # avoid inf/nan
     e = np.exp(np.array(w) / t)
     dist = e / np.sum(e)
     return dist